Neo4j Live: Entity Architecture for Efficient RAG on Graphs

all right um welcome everybody good morning good evening good afternoon uh and thank you for for joining our Neo

for live NE forj live today um at end of January so we are already you know close close to finishing first month of 2025

so it's uh it's progressing as as always um so I'm very very happy to to have you here joining us today and I'm super happy to have enina adamchik to today

joining me to talk about entity architecture for efficient rack on graphs uh Hi enena how are you doing today hi Alex um I'm doing fine a little

bit nervous of course I'm very happy to be here thank you for having me um yeah I'm very happy to no no need to be no show you what I what I brought today

yeah no need to be nervous for sure we are we are friendly friendly bunch of people um so um as always always um if you are watching this live um and you

want to reach out to us um via VIA chat please do so for any questions any any comments any feedback always be be uh be interesting to hear what you what you think of uh of today's episode of you

know the episode in general if you have any questions for me or for arena hopefully more for arena then uh then also please uh please type type it in chat and uh generally

let us know uh what you think so um Ena before we dive into the topic um uh which which you wrote a a little series of of Articles blog posts about it uh

end of last year I just uh wanted to give you the chance to let everybody know who you are what you do give uh give people a little background information on um on Arena and uh and

you know your your you know walk walk what you do currently what you excited about and and why why graph yes um yeah thanks again for

invitation um my name is Arena I'm um from accenta um ASG Center of advanced AI um originally I'm from Estonia but

now I live in Germany many years um I came here to do my PhD in physics in

2010 and um as I was already um with it I started working as a process engineer um in an very different

industry um I was working in semiconductor nanot technology field uh

almost 5 years um and after this time I I I open I discovered for myself data science and machine learning and I understood that this is something what I

want to proceed with and work on and um since then I work for consultancy uh in accenta I'm a bit over

three years now um uh I'm employed as a full stack LM developer basically J engineer and all the projects I'm doing

we are um anyhow somehow connected with geni starting from development of small

PCS um and like geni Solutions up to uh development of um strategies how to implement uh geni at the Enterprise

scale so I do quite a lot of different stuff and beginning of last year I was working um on some interesting PC and um

I um something didn't work well as I wanted and then I thought that maybe I can try graph and um I started exploring

this topic and um I came across to the very new feature by that time new for J that I could now have embeddings in my

graph and I have Vector indexes and then I started exploring this topic and since then um graph U became a big part of my

work and I'm very passionate about developing it um further um yeah and

um the PC I was working on uh basically led to a development of this methodology which I'm going to present you today um

I've built several PCS already using this um technique and um yeah I'm I'm happy to present it to you and um answer

all questions um as much as we can do it in one year one hour here um very cool um so that that's that's uh that's that's cool so you you

you have some background in in an ni and I have to ask this how much has your how much has your work changed or has your your work life changed in well in the

last one and a half or so years with everything around AI kind of like exploded this whole geni thing all of a sudden was everywhere and uh

exactly that that's the case um it's really H changed everything because by

chance I got to the geni project um very early um like November 2022 was the three and a half GPT was out and uh I

think after like around two months I got to do my first geni project we were working with Microsoft directly even uh

on this project and um and this as I started it I understood that this is something what I want to proceed with and

starting um from this project now on I'm still working on geni and um this is so such a fast developing field and there

are so many interesting things happening all the time and proceed happening and um I also part of it and I'm so proud of

it and I'm very happy to um to do it yeah the life has changed I was a data scientist before now I'm called Jia expert

yeah so yeah interesting stuff happening definitely and it's great that you that you had the possibility to make the change that this is now more more more

exciting more fun uh for you to um to um to proceed so that's that's really great to see that there is there's that and that makes it you know for you more more

relevant more you know more more exciting to to pursue that and then you you you said you you something didn't quite work and then you you discovered

graph I guess we we we we come to that so was that some kind of natural kind of almost like a natural saue where he said okay something is not working I need to research some things and then graph

popped up or did you think okay maybe a graph solution could be could be helpful here or have you have you you know looked around that the you know gen

graph rack this this kind of stuff that that was happening um towards um I think early last year middle middle of last year I guess it came up did you did you did you follow that and then thought

okay yeah actually this is this is something I I need to look at yes yes exactly that was exactly the case um as I said um I started working I think it

was February or March last year um on a project um where I had to I have been given a task uh

to develop a solution which would generate change management plans using gen it was an internal project um and

and um I have been given quite a lot of very different um documentation on it um quite a lot of different formats U

PowerPoint slides PDFs um I don't know everything web pages whatever and um I started working on it it um I I I

started with with classical rack I built up a vector database and started testing it and I understood that nothing kind of works well because um it's um it's a

little bit of mess if you do retrieval on and also how to put it all together how to you know organize it in a way that you can really utilize especially

in my case I had um I had to build tables I had to generate plants in a very structured way and um I really bumped in the problem and I was sitting

and thinking okay what shall I do now it doesn't work the only classical um solution didn't work out for me for this use case and then I started looking

yeah and I I bumped um first of all on Microsoft graph Rock article I have it in my slides also and um I started

reading that neuj um by by this time um had a very very new feature um allowing you to have a vector index on the graph

and I thought that um I mean um it was an more or less R&D project I had the possibility to try things out I asked my management they said okay go for it and

uh this is how it all started right and but the story um the story I have in my slides um which I'm going to tell you

how it's actually all evolved right so because um I did start it with Microsoft Tru I tried it out and um then the fixed

entity architect was born yeah which is a bit different thing yes super cool now I didn't want to I didn't want to take anything away from from your later slide

so I guess that's that's um that's where where where we where we can go now um as I said in the beginning if you have any questions any any comments then let us

know in chat I will share the links you you've written down um two blog posts uh I think they are already in the video description on on YouTube so if you if you want to want to read up on on on the

stuff that inen wrote before then you can do that already um but um I'll I'll give you the floor now let's let's talk

about entity architecture for uh efficient uh r or as you as you say here rock on graph fixed entity architecture um yes so um I think I've

did quite a lot of introduction already so we can um uh Deep dive directly to the slides

um um for the reason of uh um data privacy I cannot show you directly what I was working on so um this is something

very very simple just to explain the concept how I did it um I took the very the very famous sentence uh to explain

the graph which you will find everywh internet H the sentence is about Albert Einstein and um it sounds like Albert

Einstein developed theory of relativity which has re re revolutionized theoretical physics and astronomy right so and if you take this sentence you can

easily build this graph which you see on the screen which has four nodes and three edges

so um I think I don't need to describe more than that and as I already mentioned uh um April

2024 there is a very um very very famous article from coming from the Microsoft research group it's from local to Global

a graph rack approach to query Focus summarization and um Thomas banik has written um a very nice review article on

it on medium and you can read it and uh if you didn't read this article I very much recommend

so by that time if you said graph rck everybody have been thinking about Microsoft graph Rock so and that's why

um I did this presentation um I also presented it an Accenture once um I I started with Microsoft track just to explain people that people you see what

I did is something very different and I first want to show you um the difference between the two St two things right so what Microsoft uh did they took a very

large text corpor um from different domains a lot of text and uh we started extracting entities from the text using

llms uh then they started summarize the entities uh to group them in the community write the summaries of a

community and then you could do graph on the graph rack uh on these communities so um this technique is very cool it's

maybe I don't know maybe some like Google to uh kind of um thing what you can build using this

technique uh but this technique has quite a lot of disadvantages in a sense like if you work for The Big Industry this is very good for research

but for me unfortunately it didn't work out and I explain why um Thomas banik wrote uh in his review article is that

even for such a small group 13,000 notes we would require 29,000 llm calls right oh wow

so that's the problem so th this is where the problem started um and I I I started with this technique I I tell you very honestly I did two graphs using

this I don't want to know how much my employer paid for po by that time because I had a lot of

documents yes um but this is not even all um when the graph was already ahe had over

100,000 notes and even Moes and I sat and looked at this graph and I didn't know what I have to do with it because I had so many duplicates and

it was such a mess and I had a lot of space Parts data and incomplete information everything it was everything there I started looking for um articles

um I I saw that a lot of researchers are now writing articles how to D duplicate the things right so so basically uh people create garbage and then try to

get rid of it um yeah so I sat down and thought no that I mean maybe it it is the case uh

for some cases very have a lot of text like Microsoft did and no doubt but my use case was different it was smaller

and I wanted to have the control there so I I sat down and thought okay how can I do it so that I by Design do not have

any garbage right in my craft so by Design so that was my um idea and um this is how I started thinking um I have

too many duplicates in my like I I it was so messy I I could not I didn't know where to start right so it was too expensive it was definitely too

expensive I I was thinking like if I now come to my client and say oh dear client I will you know implement the state of art technology for you and you will be

very happy but you will need to pay like thousands of uh hundreds of uh Euros just for Crea creation of a database and they would say are you not

so that won't work so it was too expensive it was too complex too clattered it was real mess yeah and I really wanted to have control over my graph and I wanted to have a coste

effective solution because in consultancy if you want to be if you want to sell something um you really need to be cost effective you need to be

effective in all means right so so this is how the fixed entity architecture was born um yeah yeah

so it was a change management domain I'm not an expert in change management but um I had two business analysts and a team who were right so I started

thinking that I want to have it controllable so I need I need a fishbone of my domain to be fixed right uh and

how can I do it um I mean that's the funny thing this is the philosophy behind it many people who do

something uh do not know that they actually know their antology right MH so I took this business analyst I I put them together in one room and I

explained them what antology is and um I think I I I tried to explain um a little bit on a

very simple example I said look um I'm a person I build Solutions um this Sol Solutions we use

llm and the solutions bring money to my company which hires me and I work for it right and I also link to I like to drink coffee and the coffee grows in Plants

right y but the thing is that to build the solution I don't need coffee plants right that's clear yeah but if if if we are like um honest to ourselves but I

don't also need coffee so so basically yeah basically this the fishbone of my microw world of my domain right and I I do know it just sit and

think and you do know you can come up with it it's absolutely easy yeah right so um yeah often often times ontology as a word sounds very Grand and very

complicated and it can be but it can also be you know if you break it down to start somewhere and and and start simple and have a have a have a starting point that

is you know don't be afraid of of making it too easy that's that's fine yes exactly and

and I would even recommend to start easy and that's the key because um the less um onology layer the fixed layer I I

called it fixed layer the first layer um the less things you have Yeah the more control you have over it you can add it later but um you know be more structured

have this control so so we built this ontology as a first layer I have put the descriptions of every entity inside the note I it was

all embedded edges notes everything embedded with descriptions in descriptions I also needed um um for the

later thing which I will explain um I just wanted to um avoid um a symmetry of the dot product I mean if you if you do the similarity search right if you do

the similarity you don't want to have a symmetry in your quiry and in your something what you compare with right so so more or less you need to have at least one

length of a text so that's why it's very good to have descriptions on top and one trick which I can I can I can tell you also

um I have called all my first layer entities entity and all the names of a label of entities put to the to the properties uh because of one reason

because um if you build later on the vector index you um can build it only on one internal label it means that if I

want to search over the whole layer I need all notes be called like entity or whatever but one name right so that that

was the trick so so so the first as the first layer was ready I started putting the second layer and and the second layer actually my documentation what I

did I did the standard uh chunking embedding all the standard stuff would I would do for for for every Vector database right and now I needed to

connect these two layers um and how I did it um I just wrote a cipher query which said take all my entities all the

descriptions on all embeddings I had already on it and match to all these chunks and if you have a threshold of

similarity higher than some um some number uh connect and that's it and this is how the two layers we just glue to each

other and the and the very nice thing is that you can you can disconnect them at any time and if you add the documents you can again chunk do the same stuff

and connect using one SI of quiry so it's so scalable it's absolutely unbelievably scalable and

easy to do right so so this is how every time you add a new document you would run a new similarity Search and say okay this is this is the new value and then you realize okay yeah add add

relationships here or maybe yeah okay yes so so you basically what you do you you you have your first layer which is your ontology which you define yourself you have full control on on it and what

you do is you just you know start um adding documents as a second layer just by cosine similarity a DOT product a simple mathematical formula which you

can just include in your Cipher code and that's it you don't need much more and um remember we didn't use any llm call yet we have a graph

ready so it's absolutely for free in this sense right so we have a graph in place um later on I actually so

so the first PC I did with two layers and my first article um shows these two layers uh but later on I also added the third

layer uh which I also I mean my idea was uh already that I Want to Be llm Free um as much as possible of course you can do it using llm but what I took I took

Spacey which is an open- Source um NLP library and um what I did is just um I extracted entities from the

chunks and then I had my fixed entity Leia which is my antology which was you know created by the domain knowledge people and um then I had my documents

layer as I already discussed these are my chunks connected to to this layer by this cosign similarity and I also had the spaces

layer spaces entity layer extracted from the chunks right and this all together was connected with this cosign similarity connection trick right so now

I had three layers um this is by the way the example of the real stuff uh here you can see um these

purple ones or I don't know what are the pink purple um these are chunks of your documents so this is one document it starts here and then every chunk is

connected with next and previous Ed together and this it kind of you know contains like it's like a chain right and um these orange ones this is my first layer this is the layer of

ontology and you see this connection here I have a property here called cosign similarity and this is something what happens absolutely automatically if

I connect the layers together I always uh calculate the cosine similarity and if the cosign similarity was higher than my threshold which I has hard C it in

The Cypher Cipher quiry it was connected so so this Edge was um basically produced automatically yes and this

coant similarity is always there and what I did later on um I show you I was actually traversing the

layers by and ordering by this cosign similarity so I was uh basically uh looking for the stuff which is connected

with the shortest path or the highest uh value of cosine similarity right so uh this is the example of uh fre layers

architecture so um the orange ones are um the fixed entities then we have chunks of our documents and then we have our space entities extracted from each

chunk sometimes we overlap as you see which is a nice feature which is actually was intention to have it so right so at the end um all these relates

to edges are again having the cosine similarity property which means it's all um connected by um cosign

similarity method right so now we have three so and this is what I I mentioned in at the beginning um I always have

um one color for a layer because um I call um The Entity entity or I call documents documents or I call space

entities space entities and all real labels all real names or properties I put everything to the property because it gives me the possibility to build the

vector index on one layer all over the layer so if I will search for I will search um I have an index applied to the layer right and this is a very important

Point uh which you have to think about so as I did it all I realized that I'm I'm I'm so happy because I could do um hybrid plus search on it so I could

combine a lot of different things I could combine Vector index search I could combine full text search I was

writing I called it smart um search uh um AC quiries um so I was um I was not only using Vector index

I was also doing Dot product just an inquiry directly and um traversing through the layers and I would say so

the fantasy how you do the search on this system has no limit it's only your like your personal limit of your fantasy

because you can have you can create so many different TW twixs and tricks with all this combination um that you would never ever be capable of doing that

using some standard um Vector database like pH and like open source first um I don't want to do any advertisement for

commercial ones so right so this is a I don't I don't go to the detail but this is one example how can you tweak your uh

search uh what I did is I I took the quiry and I extracted the entities from this query from this

question um then I extracted a subr graph on all notes on of all three levels using these entities of course

yeah so then I took only fixed entity layer from the sub graph and I uh took um and then I

applied some keyword uh indexes on it so so and then you get some answer so uh what what I only want to show is that um

you can really do a lot of different stuff depending on your use case depending on what you want to achieve I mean text to Cipher do not work here I

can tell you these Cipher quiries are extremely long and complex sometimes but they do a lot of very smart stuff which

really works right so okay yeah um yeah and just coming back to Microsoft graph rocket this is something what my colleagues were asking me are you doing

Microsoft graphic no no and this is very different stuff and this is something um that you really have to distinguish so Microsoft craft rck is suitable for

large text um data sets like I said we took a large text corpora um it's adaptable to various domains so it it

actually do not doesn't it's not important what you have for you can take a text from everywhere from the whole internet and do it right okay but you

come with a high computational cost right because you have a very heavy Reliance on LMS for entity extraction

and summarization right so this stuff is expensive and um the fixed entity architecture um has a very low computational cost so basically if you

use Spacey even um what I did I did the whole graph of fre I didn't use almost no llm uh was involved in creation of a

graph itself right so um and um yeah so but you working the well defined narrow domain so so this is which is most often

the case by the way so you always have like a use case you have a domain and you have experts in this domain yeah and um yeah it requires pror of domain

specific um entities of the me of I mean you really need somebody who will create this first layer for you you with descriptions and then you can do a lot

of stuff with it right yeah okay so here I wanted just quickly to show that um having the fixed entity uh architecture

graph you can do most stuff so basically everybody wants to have reinforcement learning but if you have if you use like a GPT model which is a general um

General usage foundational model you I mean you can find unit but you don't want to do it because because it's quite expensive yeah so um so so basically you

can't uh and what I was thinking is that using the graph you can have a kind of a work around to make some reinforcement learning so imagine that you have your

graph and and this graph um works just like a r um retrieval for you for the rck system and then you user gives you

feedback like it's a good one or it's a bad one and what you can do you you could actually write this feedback back to the graph you could create the fourth

level um and just using the same idea of um cosine similarity attachment of the layers you could do it and next time you could retrieve um the best uh you could

put weights on it right you can put weights if if your user put some scores it's even better so um what you could do is basically you could retrieve also um

the most simp once um traversing from this feedback layer so so you really as I said um your fantasy is only Del limit if you have this graph you can do quite

a lot of stuff all right and this is the the last um last um uh PC what I've built for one

of our clients um I was actually quite expir in inspired uh by the connected data London um last um year in December

uh quite a lot of uh discussions were about semantic layer and I started exploring this topic so um the thing is

that if you come to the client most of the time the client has a lot of data a lot of data and if you say that okay client now we take all your data and put

it to the graph you say oh no it's too expensive it's too big it's like you need people who can do it no way right so so what you can do again work around

I like work around you can put a semantic layer on top of of whatever data the client

already have right so the semantic layer is basically graph and what I and what I did is I built this graph again using the fixed entity

architecture so I had a data which was very much numeric there was no text whatsoever there were the text was only the names of the columns and the table

right so everything what I had so um so I created the layer of data Shima so this layer was basically only the column the table names and the column names and

they were connected to each other right and then I created again the the my fishbone layer which is a ontology

layer G giving some descriptions and so on and um because this layer it had no descriptions so like almost no text what

I did I just again used my cosign similarity match trick and connected these two layers and yeah it worked out

I started basically retrieving the information on my data sending it to uh llm asking to create SQL the SQL was

created then the SQL then the data was retrieved from the relational database and the answer was summarized and given back to the to to the to the user so

what was done user WR the quiry and I vectorize the quiry and I have the vector index on each of these layers here right yeah so what I do

next using this layer I match and extract most similar ontology entities to the user quiry so I match this quiry

to here yeah next I take this layer I imine extract more similar table column names to the user quiry

next what I do I do the traversal I I say find the shortest path with two or three hops from one layer to another and extract me the connections

and and build me this path and this is already a real semantic reasoning so people who do rdf can tell me that there is no semantic reasoning in property

graph I don't I disagree I do it so um yeah so I extract I extract the short shortest paths with

two to three Hops uh from the antology layer down to the Shima layer and I do the same from top to the bottom oh from bottom to the top right and what I do

next I do the union this is just in cyer I do the union and I extract it all and I give it to L and it works very cool

actually great yeah so this is the latest Discovery so now I I come to to the end of my presentation uh very

quickly summarized fixed entity approach it has two or three layer uh knowledge craft which allows you to avoid

duplication uh of the entities by Design um you have um documents and named entities are added and connected using cosign

similarity um you you would have a possibility of kind of reinforcement learning using the graph right so um which you would not be

able to do with u GPT models or any kind of General usage models which do not belong to you right no yeah um you can build a very flexible hybrid search

doing quite a lot of stuff uh as I said your fantasy is your limit and um yeah you can connect different knowledge domains as well um if you build build these fixed entity graphs you can

connect them together and Traverse again so um this approach worked well for me because for the client I I was able to

create something without hundreds of thousand llm calls so almost free let's say um so the graph is ready very quickly it's very scalable you can add

as much documents as you wish and you can build your search quaries also as you wish and this is a real cool stuff

and um I still in the process of development this approach further and yeah thank you for your attention and let's stay connected let's stay in touch

cool thank you very much enena that was that was very exciting very interesting um I have um a couple of questions so obviously if anybody else has any questions please please use the chat

function to to write your questions um farida asks uh on LinkedIn um if I want to have some specific entities in hand and a knowledge graph in hand if I want

to extract information relevant to the extracted entities from the knowledge graph what should be the best way so she says should we perform semantic search

only or hybrid search or or what what would you recommend on yeah I I would I would recommend um I mean my experience say that um if you have embeddings build

the vector index it works better than the uh keyword search it works it works definitely better but what I always did

I always connected them I always had a union um of these two worlds um and that's quite easy to figure out if it

works or not for your use case basically but try both yeah I would say but start with Vector search it it's really good it it works very well but what I was

also doing uh sometimes was just coding the dot product inside the query and doing the dot product on the user quiry first extracting the sub graph and then

applying the indexes that's also um like a smart search way how you can do it but you need to uh experiment for sure yeah

good good feedback um thank you for that and the other question here is from amot and uh they ask if uh the slides would

be available some somehow somewhere can we share them with the audience later on

um I I need to come back uh to you with the answer um not sure if I may no no problem if not the the uh the summary as

I said initially is also in the in in the article you wrote on medium so there there's there's lots lots in there as well that is um also good that's true that's true check the

article at least the the most important stuff um I this infographics um of the free layers I do have yeah yeah it's in

the it's in the video description on YouTube I think it also is on LinkedIn in the video description um and I put it just in chat one more time um if you

like I said any other question please free to type away um if not I think do do we have a demo now

inen yes um yeah so um it's not so exciting maybe to see the the ready product because you don't see what's

happening behind but but yeah this is um the very first um thing what I did um with this technology so it was as I said

already it was a change management um project so what uh the task was to create a change management plan

um I had I had like guidelines how to create the plan and there are a lot of different things different modules and different activities belonging to the

modules so quite a lot of stuff and um what the graph is doing behind it's actually it's filtering it it traversing the right modules right

activities um looks at the user quiry find is the most matching one uh uses the filtering of a similarity property on the edge what I showed you but this

all you will not see here but this is how this thing works so what I do here is um I select um a type of a program my

select duration budget this is all just a Sandbox UI it was created by me just to show the concept so so you see that

um this is a fully generated graph here at the right uh right hand side um it's a reference of a documents which were

extracted from the graph okay and yeah so you can basically um you know write some some some iteration of the graph

what you want to change and then you can iterate on it and then this is something what I just showed how the iteration was done um yeah so you don't you cannot see

probably much but this is um this is basically a working graph rack solution based on fixed entity

architecture what you see here right so so yeah that I cannot show you much more no that's fine sorry for

that no just I think today today's session is a little more more uh you know not not not saying it it's it's generally theoretical but gives you it's more of a of a uh I would say inspiration to to

everybody out there working with graph rack to to to see you know what's possible what's out there different concepts different ways of of um of working with it I I'm definitely

intrigued I think it's a it's it's a cool um cool idea and it also takes away the like you said in the beginning this this D duplication process this The

Entity recognition step a little bit um is is is really hard and it's really tricky sometimes to get to get to a right Baseline the ontology the semantic

layer all of that is if if you Outsource all of that Tod llm it makes lots of mistakes it is not really your your domain it's not really relevant to you

and it and it creates noise and if you if you can take away that that problem and with your with your semantic um sorry with your your similarity um

calculations and then the linking together I think that's that's a great step and then you um you you you get better results in the end and it's it's good to try it out so if you um you know if you're working on

a graph R problem and you you feel like oh this is this is um this is a cool solution let us know uh try it give give it give us give us a you know let us know if you did it let enena know if you

if you tried it out and if this helps absolutely I'm very interested and yes is please write me in LinkedIn um and share your your results and I'm I'm

really very interested to see how it works for different use cases as well right so as I said let's stay connected

you will find me in LinkedIn and U I'm very happy to hear what was done and maybe also the new ideas because we evolve all the time like with this

semantic layer now uh so so you you you have this concept and you can take it and go and do something else with it so absolutely yeah that's funny right you

with two layers then you added a third layer with space now now you almost added a fourth layer with this um you know feedback loop who knows at some at some point we might have five six seven

layers but it's all it's all connected and that's the beauty of the graph as well you can expand it you can add to it you can add more um more more knowledge

to it and it becomes even even better so it it's it's not it's not a problem it it is it is even enriching the data more and gives you better results in the end so it's really

good cool um I think with that um we we are the the end of today's um presentation today's session uh before

uh I let you all go I'd like to you know let you know or highlight upcoming session so um on the 28th of January so in a next Tuesday I'm having Quinton

rosil joined me about um taming llm hallucinations for medical Q&A so maybe even you know it's a it's a good good segueway here because of uh of of the

problem of U the publication and uh and noise so will be interesting to to see what what he's gonna g to share about

his story about his journey uh and how to um make a a medical Q&A chatbot um system we still run our new forj

developer survey so if you are uh interested in giving us feedback then uh please let us let us know uh through the link and through the the QR code you can

uh you can follow U the step and um if you want to know more if you want to continue now uh with a with a graph rag application or with something else in in the ne of JW then here are a couple of

links that are interesting uh and helpful we hope graph Academy to learn stuff Neo forj Ora free is our offering for a graph database as a service so you don't have to install anything it just

works um and more stuff from the NJ Community is on our community page we have a Discord server we have a community forum and and much more so let

us know and obviously if you're watching this on YouTube um if you liked it give it a thumbs up let us let us know that this is this is worthwhile your time if

you have any feedback any comments put it in the video um the comments in the chat comments and um yeah for that uh

thank you already and with that end end of the session today thank you very much enena again for presenting today for writing the blog post but obviously also for taking the time today and showing us

through the uh interesting approach of your graph rack architecture and uh thank you everybody for watching have a good rest of the

week thank you thank you Alex